annual report (year two) on the habitat evolution mapping project for … · 2011-07-15 · annual...
TRANSCRIPT
Annual Report (Year Two) on theHabitat Evolution Mapping Project
for theSouth Bay Salt Pond Restoration Project
Submitted by:
Brian FulfrostProject Manager
Habitat Evolution Mapping ProjectBrian Fulfrost and Associates
July 6th, 2011
July 6th, 2011
Annual Report (Year Two) on theHabitat Evolution Mapping Project
for theSouth Bay Salt Pond Restoration Project
Submitted by
Brian Fulfrost, Project Manager
South Bay Salt Pond Restoration – Habitat Evolution Mapping Project
Brian Fulfrost and Associates
56 Rio Vista Ave
Oakland, CA 94611
831-566-7686
Summary of Work Completed
This report outlines the work conducted during Year Two (April 2010 – April 2011) of the Habitat
Evolution Mapping Project (HEMP). In Year Two (2010) the HEMP team successfully utilized an
improved version of the habitat model to map over 15,000 acres of marsh habitats and mudflats in the
south bay of San Francisco. We spent considerable amount of time improving the habitat model's
overall accuracy (already at 80% attribute accuracy) from Year One and providing the necessary data
for tracking changes to habitats within the study area, including ponds already breached for restoration
(see Figures 1,2 and 3). Improvements in Year Two focused on increased efficiency in ground truthing,
changes to the mix of habitats being mapped, and tweaks to image classification methods. Although
there was some delays in receiving the Ikonos imagery within specification, there were no significant
delays from reporting, permitting or equipment that were previously encountered during Year One. As
a result, there were additional improvements in overall project efficiency, Ikonos imagery acquisition
timing, the effectiveness of our ground truthing, as well as in our final Year Two model results.
Status of Mapping Products
Habitat Classes
The list of habitats being classified in the model was further refined in Year Two, based on a number of
factors. These include the effectiveness of a given habitat class during a model run, the spectral
similarity (or difference) between a given habitat and another, the image characteristics of certain
vegetation associations or habitats, , and ultimately the importance to overall project goals. Changes to
the list of habitats being mapped were vetted in consultation the South Bay Salt Pond Restoration
(SBSP) Project Management Team (PMT) and US Fish and Wildlife Service (USFWS) biologists at
Don Edwards National Wildlife Refuge.
Biofilm
After reviewing the results of our initial habitat classifications during Year Two, it became clear that
there were “out of place” biotic signatures appearing on mudflats in both the 2010 and 2009 imagery
(Type I errors of other habitats not found on mudflats, notably Gumplant). These areas either did not
normally contain macrophytic plant species (on the mudflats themselves) or plant species of this type
(high marsh plants vs. low marsh plants). Dr. John Takekawa (USGS) and other researchers identified
that the likely source of these signatures was “biofilm” occurring on the surface of tidal mudflats. This
1 Year Two Annual Report
biofilm might potentially serves as an important source of food for bird populations important to the
restoration effort. As a result, the HEMP team and the SBSP PMT decided to add biofilm to the list of
priority habitats to be mapped.
We obtained ground truthed locations of biofilm in order to generate training sites to map the extent
and distribution of these types of mudflats throughout the study area. The result of adding biofilm to
our list of habitats classes, include:
• a validated map of biofilm distribution on mudflats throughout the study area (within breached
ponds, along sloughs and even within marshes); and
• a significant improvement in overall model accuracy especially within breached ponds.
After subsequent conversations with the USGS team studying biofilm at Dumbarton shoals, it was clear
that (a) the lack of water on top of and exposure of mudflats to the sun resulted in potentially higher
probability of biofilm; and (b) there were a range of type of biofilms which could be descriptively
characterized by density. We are not confident that our ground truthing locations adequately represent
this variability and therefore we can not be certain if the model is mapping the complete distribution of
biofilm.
By the end of Year Two, the model contain a total of 25 unique “habitats” (see Table 1 and Table 2).
These are the various biotic (e.g.“Pickleweed”) and abiotic (e.g. “mud”) land cover classes that are
mapped as part of the project. The biotic habitat classes include: 10 salt marsh, 4 brackish, and 3
freshwater vegetation associations. Although the focus of the project is on vegetation, and vegetation
associations (e.g. Pickleweed /- Cordgrass), the term “habitat” is being used here to refer to both biotic
and/or abiotic land cover classifications.
2 Year Two Annual Report
Table 1: Marsh Habitat Classifications
Salt Marsh
Community
Salt Marsh
Vegetation
Brackish Marsh
Vegetation
Freshwater Marsh
Vegetation
low
Pickleweed (annual) Alkali Bulrush BulrushCordgrass Alkali Bulrush /-
Pickleweed
Cattail
low - mid Pickleweed /- Cordgrass Alkali Bulrush /-
Pepperweed
Bulrush /- Cattail
mid
Pickleweed /- Saltgrass Alkali Bulrush /-
SpearscalePepperweedPickleweed /- Pepperweed
Pickleweed /- Jaumea
mid - high
Pickleweed (perennial)
Pepperweed
high Alkali HeathGumplant /- Pickleweed
Table 2: Non-Vegetated and Upland Habitat Classifications
Non-Vegetated Upland (and ecotone)
Mudflat Alkali GrassesMudflat with Biofilm MustardAlgae SaltgrassWrack Alkali HeathBare Earth PepperweedWater
3 Year Two Annual Report
Ground Truthing
In Year Two, we surveyed 172 locations using one of our standardized ground truthing protocols and
noted the vegetation and other land cover characteristics at an additional 331 locations (see Map 1).
David Thomson, the lead biologist on the HEMP tam, conducted the majority of our ground truthing
surveys. These ground truthing datasets served a number of functions:
• evaluating the number and type of habitats being mapped,
• assisting with the qualitative review of model results,
• identifying new or modified training sites for various versions of the model, and
• final model validation.
As in the previous year, we continued to utilize the Trimble Yuma (with Trimble Terrasync for data
collection) with a Trimble ProXT external antenna for obtaining sub-meter GPS locations of our
ground truthing sites. During Year Two, modifications were made to the Terrasync “data dictionaries”
(i.e. our digital field survey forms) as necessary and as suggested by our lead biologist, David
Thomson. Significant time was also spent in cleaning and combining our ground truthing datasets so
they could be used more efficiently. We used ground truthing surveys to assist with developing new or
modified trainings sites for the habitat model as well as for model validation.
Rapid Assessments
Between June and August, we conducted 69 additional rapid assessments to assist with supplementing
our knowledge of vegetation associations and phenological variability throughout the study area.
Model Calibration
Between June and September, we also conducted 30 ground truthing surveys (a simplified form of
Rapid Assessments) to assist with calibrating the model results.
Validation
Unlike Year One, we were successful in obtaining a stratified random sample of 73 habitats on the
ground for statistical validation of the model. In order to identify the locations for our validation
surveys, we generated a random sample of 200 points stratified (by area) into the following categories:
salt marsh, brackish marsh, fresh marsh, and managed marsh. Our validation focused ground truthing
4 Year Two Annual Report
occurred between October 2010 and November 2010. Unfortunately, due to time and weather changes,
we were only able to survey a little more one-third of these 200 locations. As a result, we supplemented
our validation sample with ground truthing surveys obtained from our rapid assessments and model
calibration during Year Two.
The table and chart below, which do not include supplementary data and exclude data collected on
levee tops, help to characterize the type and variability of dominant vegetation associations among salt,
brackish and fresh marshes within the study area. Table 3 below lists the number of occurrence and
type of dominant vegetation associations obtained from our ground truthing (excluding our
supplementary ground truthing dataset) . Chart 1 proportionally divides the various types of vegetation
associated with Pickleweed (perennial) obtained from our ground truthing in Year One and Year Two.
5 Year Two Annual Report
Table 3: Number of Occurrences of Dominant Vegetation Associations (Year One and Year Two)
Chart 1: Pickleweed Associations (Year One and Year Two)
We achieved approximately 80% attribute accuracy using a total of 172 points for our preliminary
validation drawn from these three ground truthing surveys. Once we have filtered and cleaned our
supplementary ground truthing surveys, we will use these additional locations to expand our final
validation sample size for Year Two (and for Year One).
6 Year Two Annual Report
Map 1: Year Two Ground Truthing Locations
Ikonos Imagery Acquisition and Processing
The 1 meter Ikonos imagery was acquired on July 4th, 2010 (Year One date was June 23rd, 2009) at 2.3
feet above MLLW, allowing for maximum exposure of vegetation and mudflats as well as reduced
moisture content in the vegetation. The image was acquired on the second attempt (out of a possible 3)
since the first image capture contained too much cloud cover and was out of specification.
The HEMP team selects 3 days in June or July, when tidal marsh vegetation is near peak growth. (In
order to select specific days the project pays a “tasking fee” to secure the satellite on days the company
say its available.) The IKONOS satellite passes over the study area around noon to minimize shadows
and glint from the sun. Our optimal acquisition days are when the lower-low tide occurs 1-hour before
noon and the predicted tidal depth is less than 1 foot above ( or below) the station’s MLLW datum. We
optimally want one hour after lower-low tide because there is a delay in the tidal cycle from North to
South “up-estuary”, which is roughly ½-hour between the San Mateo Bridge and Alviso. As a result,
during the hour following an ebb tide the tidal depths across the study area will be more equivalent
because northern areas will be in slack tide (i.e. not changing) for at least an hour while the southern
portion reaches low slack tide. However, since days at which the Ikonos satellite will overpass do not
always match with these optimal dates, we had to loosen our criteria to 1 hour before or after lower-low
tide and a predicted tidal depth of less then 1 meter.
The HEMP team, in consultation with the SBSP PMT, also considered switching to the higher
resolution GeoEye1 satellite for Year Two but stayed with Ikonos to ensure consistency for the time
period of the study. The GeoEye1 also lacks operational bi-directionality and therefore it is more
difficult to capture the study area in a single pass.
The Ikonos imagery was obtained in 3 “snapshots” from a a single pass over the study area. In Year
Two, the Ikonos imagery was delivered in two orthorectified image formats: (1) as a single 1 meter
pan-sharpened image mosaic (of the 3 “snapshots”) that have been radiometrically processed for image
clarity; and (2) as 3 individual “strips” containing the “raw” Digital Numbers (DN) for both the 4 meter
multispectral and 1 meter panchromatic bands. The pan-sharpened image mosaic is used for all general
mapping needs as well as our crucial qualitative model review. The “raw” data, contains unmodified
values which have not been radiometrically corrected (e.g. DRA), pan-sharpened, or mosaiced. As a
result these image strips contain non altered (or “raw”) image values (also called “digital numbers” or
DN) and are consequently used for image normalization, habitat classification and change analyses.
7 Year Two Annual Report
Unfortunately, the imagery was delivered out of specification with regards to positional accuracy and
as a result we didn't receive the final satellite imagery (within 2 - 4 meters) until late August 2010.
Image processing in Year Two focused on (1) developing standardized methods for normalizing and
correcting the Ikonos imagery for applying the model to allow for consistency of mapping over time as
well as for performing change analyses; and (2) improvements to the accuracy of spectral habitat
model developed in Year One. Erdas Imagine 9.3 continued to be used for all image processing, habitat
model refinements, and image classification.
Image Masking
Once we received the final Ikonos imagery for Year Two, we utilized the inclusion/exclusion mask to
generate imagery that include only the area within the project boundary actually being mapped. This
assist in both qualitative model interpretation by reduction the area b being classified as well as
qualitative image classification by only including the elements of the baylands that the spectral model
is classifying.
The HEMP team spent some time editing the boundaries of our image “mask” to better comply with a
actual marsh and pond boundaries. This mask has 3 versions, all of which cover the exact same area:
(1) an inclusion/exclusion mask for areas within the image that are being mapped as part of the project;
(2) the inclusion/exclusion mask divide into the following categories: salt marsh, brackish marsh, fresh
marsh, and managed marsh; and (3) the inclusion mask divided up into a number of habitat types from
the original sources of the mask (Habitat Goals, 1998; Existing Conditions – 2004; HT Harvey
Vegetation Mapping of Guadalupe Slough – 2008). Once the Bay Area Aquatic Resource Inventory
(BAARI) being developed by SFEI is released, we will utilize it to further refine this mask as needed
for Year Three.
Image Normalization
Year Two introduced additional processing requirements required for radiometric normalization of the
images in both (and subsequent) years.
The HEMP team spent a considerable amount of time in the Fall 2010 attempting to normalize the
Ikonos imagery from both Year One and Year Two. The preferred method would be to utilize an
absolute image-based atmospheric correction method such as Cosine of Sun Zenith Angle (COST) or
Dark Object Subtraction (DOS) so as to convert the at senor radiance to actual ground reflectance. An
8 Year Two Annual Report
alternative and acceptable image based atmospheric correction method would be to utilize the
standardized exoatmospheric correction (in essence “planetary reflectance” vs. ground reflectance)
developed for Landsat imagery and commonly applied to Ikonos imagery. Unfortunately, although
significant project time was spent on applying various image based atmospheric correction methods to
the Ikonos imagery, we were unsuccessful at achieving acceptable results using any of these methods.
In a effort to continue to make progress towards our goals, we utilized an alternative relative
normalization method (as opposed to absolutely correcting for atmospheric disturbance) known as
histogram matching. This is an image based correction method in which we use a “base year”, in this
case our 2009 image, and match the histograms of subsequent years to the base year in an effort to
normalize the spectral variability of images over time. The advantage of this method is it is easy to
implement. The disadvantage is this method does not account for atmospheric effects. In the end, we
matched the histograms of each 2010 Ikonos strip to the corresponding strip from 2009 resulting in
accurate model results for each year. Although our image normalization was successful, it was not
perfect. As a result, there continue to be radiometric differences between the image years that introduce
some error into the application of the model.
Supervised Classification Review and Habitat Model Refinements
Once the images were normalized, we ran the most current version of the habitat model on both the
2009 and normalized 2010 (raw) image strips (3 total for each year). As in Year One, we utilized a
“maximum likelihood” supervised classification. In the final working version of the Year Two model
we also added probabilities (in the form of statistical weights) to specific habitat training sites (e.g.
Gumplant) based on the type of error (I or II) and/or level of accuracy of a given habitat. This greatly
improved the overall accuracy of the mapping results for “difficult” habitats such as Gumplant.
ArcGIS continued to be used for our qualitative model review, digitizing training sites, as well as for
statistical validation.
Improvements to the habitat model were made iteratively, as we performed our qualitative review of
the results of a given model results and as we ground truthed additional areas and habitats. Changes to
the model resulted in the increased accuracy of individual habitats and improvements to the “mix” of
various vegetation associations being mapped. These improvements resulted from introducing training
sites that captured the adequately habitat and phenological variability in the study area and were
spectrally 'distinct' enough to be mapped as distinct habitats.
9 Year Two Annual Report
Since the habitat model consists of a series of training site(s) for each habitat class (that are used to
represent the “spectral signature” for each habitat being mapped), the primary mechanism for
implementing these changes was either to alter existing training sites or to intrude new and/or multiple
training sites for each habitat. A significant amount of project time was spent in the process of model
review and model refinement. In the end, dozens of models were developed with various numbers of
and types of habitats.
Necessary improvements to the model were identified as we obtained more data (primarily from
ground truthing) regarding the spatial, vegetative, and phenological variability of habitats. As a result
we better understood the spectral differences (or similarity) of a particular habitat. As the list of habitats
being mapped changed we discussed these changes with appropriate USFW&S staff to ensure that the
mapped habitats would still ultimately meet project goals. For example, although mapping the extent
and distribution of Iceplant (both Carpobrotus sp. and Mesembryanthemum nodiflorum ) would be
useful to the restoration effort (in in effort to understand its impact on marsh vegetation dynamics), it's
spectral similarity to Pickleweed resulted in significant Type I errors (false positive) where Iceplant
was erroneously displacing large amounts of Pickleweed . As a result, we excluded this species from
our model during Year Two, resulting in increased accuracy of other priority habitats (notably
Pickleweed).
As in previous years, the HEMP team qualitatively assessed the accuracy of a given model's results,
utilizing the following resources:
• existing ground truthing datasets (from HEMP, ISP, HT Harvey),
• Bing Maps (oblique view),
• our habitat interpretation guide,
• and expert knowledge of the study area.
Additional ground truthing surveys were obtained to calibrate the model during their qualitative
review as the image interpreter identified problematic or questionable habitat assignments. Details
about the protocols and procedures used for our qualitative review have been explained elsewhere (see
HEMP Year One Annual Report) and will be documented in our final report.
10 Year Two Annual Report
2010 Habitat Classification Results
By the end of February 2011, the HEMP team had developed a significantly improved habitat model.
We applied the final (working) Year Two habitat model to both the Year One and Year Two imagery.
Both of these habitats classifications (2009 and 2010) meet our 80% attribute accuracy requirement,
although a number of habitats require improvement (e.g. Grindelia). We will apply our final statistical
validation to all three years in Year Three, as we include changes from Year Three.
Overall, the model outcomes provide an accurate picture of the distribution and extent of habitats
within the study area at the scale outlined in our proposal (1:2400). We have highlighted important
model results from Year Two in the five figures below.
Figures 1, 2 and 3 demonstrate how the model has accurately tracked the growth of vegetation,
and possible accretion of sediment, within Pond A21 (a pre- Phase I breach) between Year One
and Year Two. Figure 1 also displays the models ability to clearly track Lepidium latiofolim
(shown in blue) and Pickleweed (shown in orange), on the left side of both images, in the
marshes around “drawbridge” between Pond A21 and A20.
Figure 4 the consistency of mapping Pickleweed (shown in orange) at Ideal Slough between
Year One and Year Two as well as the increase in wrack (or dead plants)
Figure 5 displays the success in modeling Grindelia (shown in “bubble gum” pink) dominated
high marshes along interior slough channels at the mouth of Mowry slough in both Year One
and Year Two. However, spectral and phenological differences between the years have resulted
in differences to the extent and distribution of Grindelia in the marsh.
Figure Key
• Pickleweed (perennial) is shown in orange• Pickleweed (annual) is shown in grey• Cordgrass is shown in yellow• Pepperweed in blue• Mudflats in brown• Mudflats with biofilm in purple• Wrack in red• Algae in green.
11 Year Two Annual Report
2009 2009 model 2010 2010 model
12 Year Two Annual ReportFigure 2: Growth in Pickleweed in Pond A21 (2009-2010)
2009
2010
Figure 1: Growth in Pickleweed in Pond A21 (2009-2010)
April 2009 Sept 2009 May 2010 October 2010
13 Year Two Annual Report
July 2010
Figure 3: Growth in Pickleweed in Pond A21. The 4 photos appearing on top are courtesy of Cris Benton and demonstrate additional evidence for the growth of Pickelweed (both annual and perennial). The area taken in these photos is highlighted on the Ikonos imagery and model results by a yellow box.
June 2009
14 Year Two Annual Report
2010
2009
2010
2009
Figure 5: Consistency of Habitat Mapping (pickleweed, cordgrass, pickleweed/cordgrass shown in tan, and pickleweed/jaumea shown in dark purple) in Tidal Marsh outside Pond A6
Figure 4: Gumplant (high salt marsh) mapped in both 2009 and 2010 at mouth of Mowry Slough. Spectral and phenological differences between years result in variability in results.
Spectral and Phenological Variability
Differences in the spectral values for a given habitat from year to year are a result of a number of
factors, including: image details (e.g. elevation angle), radiometric normalization, and phenological
differences (spatially and temporally). As a result, there will be slight differences in the mapping of a
given habitat in certain locations (local spatial variability) from year to year. Some of this variability
represents true phenological differences (dead vegetation one year vs. alive second year) in vegetation
and therefore cannot be accounted for through automated methods alone. In Figure 6, during 2009 the
Pepperweed (in blue) was dead and therefore is being mapped as Wrack (in red) while in 2010 it was
alive and therefore is being mapped as Pepperweed. In these cases, the model is actually performing to
specification since it is accurately mapping the land cover .
We have minimized the effect of this local spatial variability by trying to optimize acquisition details
(elevation angle closer to nadir), applying an effective radiometric normalization technique, and
including training sites that account for phenological differences in vegetation. However, we will
continue to attempt to account for this variability in Year Three by optimizing one or more of these
three techniques. In addition, we will explore additional methods to account for the variability (or
similarity) of certain habitats. These methods include mapping certain species which are spectrally
similar(e.g. Alkali Heath and Spearscale) as one habitat as well as identifying clear “thresholds of
15 Year Two Annual Report
2009
2010
Figure 6: Differences in phenology of Pepperweed in 2009 and 2010 (east of Pond A18)
change” (by area) for each habitat that account for this variability.
Limitations
• Mudflat Extent
The Ikonos image in Year Two was acquired at a time very close to Mean Lower Lower Water
(MLLW) and significantly improved upon the exposure and subsequent mapping of mudflats
(see below). However, the HEMP team met with key members of the PMT during Year Two at
which time it was decided that the issues involved with matching dates to which the Ikonos
satellite will pass over the study area during (or close to) MLLW limited the efficacy of using
the Ikonos imagery for mapping changes to mudflat extent..
• Levee Tops
In consultation with the SBSP PMT, the HEMP team will map “levee tops” using only two
land cover classes: (1) vegetated; or (2) non-vegetated. The mix of often weedy peripheral
halophytic vegetation that characterize levee tops have made it very difficult to accurately
distinguish between specific habitat types with any great deal of accuracy. Levee tops are
“linear” habitats that have a narrow width and often contain a complex mix of weedy
halophytic vegetation. In these cases, the imagery from the Ikonos was not capable of
accurately distinguishing individual habitats.
16 Year Two Annual Report
Year One Ikonos Year Two Ikonos
Remaining Improvements
• High Salt Marsh Habitats
The habitat model has successfully been able to delineate high marsh habitat. However, we have
been less able to accurately distinguish between certain habitats within these high salt marshes.
We have had success in distinguishing certain species, such as Alkali Heath, Jaumea, and to a
lesser degree Saltgrass.
On the other hand, we have had little success mapping Spearscale as a unique habitat often
because of its spectral similarity to other species like Alkali Heath, and the lack of good training
sites. Although certain locations of Alkali Heath might actually be Spearscale in some cases, in
either case, they are being accurately mapped as existing high salt marsh, and therefore can be
considered to be meeting project goals. On the other hand, one important species we continue to
try to accurately distinguish is Gumplant, a critical high marsh indicator species. Because of its
morphology, growth pattern (bushy along interior marsh slough channels), and association with
Pickleweed it has been often difficult to map Gumplant consistently throughout the study area
(there are a range of both Type I and Type II errors). The introduction of new or modified
training sites that represent this variability has often resulted in significant over mapping of
Gumplant. By the end of Year Two, Gumplant continues to be an area of focus for Habitat
refinement, although we are confident that our final results will be within specification.
Throughout the study area , high salt marshes are found in a number of locations within salt
marshes: 1) higher areas adjacent to sloughs often within the mid marsh; 2) higher 'mounded
areas' within the mid marsh (e.g. from natural sediment accretion or from man made features
such as historic levees), or 3) along levee flanks.
17 Year Two Annual Report
Presentations and Outreach
During Year Two, the HEMP team presented the ongoing results of the mapping project at a number of
conferences and to a number of affiliated organizations (SFEI and HT Harvey). Brian Fulfrost
presented the results of the mapping project at both the Bay Delta Conference held in Sacramento in
September 2010 and to the South Bay Salt Pond Restoration Science Symposium in February 2011. In
addition, the HEMP team continued ongoing discussions with project partners and other key
stakeholders on the project. Brian Fulfrost also provided guidance and oversight to the interns at the
NASA Develop program where students worked on a number of remote sensing projects related to the
restoration.
Year Three and Next Steps
In Year Three, we will finalize our habitat model using all 3 years worth of imagery, perform our final
statistical validation by focusing our ground truthing on validation, run change analyses on the final
model results, and document our work in our final report. We will also distinguish channels (with and
without water) and pannes as generated from the image analysis and perform manual edits to the
model results in problematic areas (or habitats).
Additional focus areas for improvement, include:
• atmospheric correction of imagery ( to potentially improve radiometric normalization),
• improvements to the final model results using two types of raster analyses:
(1) the integration of the tide/elevation raster developed by Gavin Archbald,
which will help remove blatant habitat mis-assignments according to the likelihood of
vegetation existing with a tidal/elevation range corresponding to a particular marsh
community (e.g. Gumplant in high salt marsh), and
(2) a raster based neighborhood analysis to reduce unnecessary habitat complexity.
18 Year Two Annual Report